2034367 - Improve performance of wasm wide-arithmetic on x86_64

Status: RESOLVED FIXED

(Core :: JavaScript: WebAssembly, enhancement, P3)

Description

[Julian Seward [:jseward]]

Our initial implementation of i64.{add,sub}128 and i64.mul_wide_{u,s} can be
slow, especially on x86_64, per numbers at [1]. We should investigate and
improve.

[1] https://github.com/WebAssembly/wide-arithmetic/issues/7#issuecomment-4292562703

Comment 1

[Julian Seward [:jseward]]

Attachment: bug2034367-wide-arith-PIP.diff-2

Reworks the x64 sequence for {M,L}MulI64WideHI64. Improves
performance for https://github.com/alexcrichton/wasm-benchmark-i128
on Intel Tiger Lake by 3.5%.

Comment 2

[Julian Seward [:jseward]]

Attachment: Bug 2034367 - Improve performance of wasm wide-arithmetic on x86_64. r=yury.

This patch:

• Reworks the x64 sequence for MacroAssembler::wasmMulI64WideHI64 and hence for
  {M,L}MulI64WideHI64. Specifically, MacroAssembler::wasmMulI64WideHI64 no
  longer tries to preserve RAX and RDX; those registers are trashed and instead
  the problem of preserving them is given to the Ion register allocator / to
  the baseline "manual" allocator.

  This makes the sequences shorter and allows removal of the "xchg"
  instruction, which (judging from Agner Fog's latency tables) is a limiting
  factor on ILP.

  No change for any non-x86_64 targets.

• Adds a foldsTo method for MWasmAddSubI128HI64. This doesn't see much
  action, but can fold out 0 + x and x + 0 cases.

Performance for https://github.com/alexcrichton/wasm-benchmark-i128 on Intel
Tiger Lake is improved by 3.5%, mostly via improved IPC.

Comment 3

[Pulsebot]

Pushed by [e-mail]:
https://github.com/mozilla-firefox/firefox/commit/8ceecb618321
https://hg.mozilla.org/integration/autoland/rev/4ac00457d12c
Improve performance of wasm wide-arithmetic on x86_64.  r=yury.

Comment 4

[amarc]

https://hg.mozilla.org/mozilla-central/rev/4ac00457d12c